基于FCOS神经网络的制动主缸内槽缺陷检测方法  

Internal Groove Defect Detection Method of Brake Master Cylinder Based on FCOS Neural Network

作　　者：王芷薇 郭斌 胡晓峰 罗哉 段林茂 WANG Zhi-wei;GUO Bin;HU Xiao-feng;LUO Zai;DUAN Lin-mao(College of Metrology and Measurement Engineering,China Jiliang University,Hangzhou,Zhejiang 310018,China;Zhejiang Key Laboratory of Advanced Manufacturing Technology,Hangzhou,Zhejiang 310058,China;Hangzhou Wolei Intelligent Technology Co.Ltd,Hangzhou,Zhejiang 310018,China)

机构地区：[1]中国计量大学计量测试工程学院,浙江杭州310018 [2]浙江省先进制造技术重点实验室,浙江杭州310058 [3]杭州沃镭智能科技股份有限公司,浙江杭州310018

出　　处：《计量学报》2021年第9期1225-1231,共7页Acta Metrologica Sinica

基　　金：国家自然科学基金(51675499);国家自然科学基金重大科研仪器研制项目(51927811);浙江省先进制造技术重点实验室开发项目(2019KF01)。

摘　　要：针对主缸内槽缺陷检测存在干扰因素复杂、检测精度低等难点,提出了一种基于全卷积单阶段神经网络(FCOS)的主缸内槽缺陷检测算法。利用特征融合金字塔网络进行特征提取并逐像素预测,得到缺陷种类,实现凹槽缺陷的自动检测。实验结果表明,FCOS网络对制动主缸内槽砂眼、划痕、振刀纹缺陷检测的平均精度均值分别为85.2%、87.5%、90.1%,精确度分别为0.98、0.89、0.95。实验结果与Mask R-CNN网络和Faster R-CNN网络的实验结果进行对比,FCOS网络具有更高的准确率,学习时长大幅度缩短,且满足实时检测要求。Aiming at the difficulties of complicated interference factors and low detection accuracy in the detection of groove defects in the main cylinder,a detection algorithm for groove defects in the main cylinder based on full convolution single stage neural network( FCOS) was proposed. FPN network was used for feature extraction and pixel by pixel prediction,and the predicted results were classified to realize automatic detection of groove defects. The experimental results show that the m AP values of FCOS network in detecting the sand hole,scratch and vibration pattern in the inner groove of the main cylinder are 85. 2%,87. 5% and 90. 1%,and the detection accuracy is 0. 98,0. 89 and 0. 95. Finally,the experimental results were compared with those of the Mask R-CNN network and Faster R-CNN network. FCOS network had higher accuracy,significantly shortened learning time and satisfied real-time detection requirements.

关 键 词：计量学 内槽缺陷检测 制动主缸 全卷积网络 FCOS 特征融合金字塔网络 

分 类 号：TB973[一般工业技术—计量学]